There is no trustworthy current leaderboard, so the most recent reliable figure is already months old.
LiveCodeBench
LiveCodeBench scores models on competitive-programming problems, but only on ones published after each model's training cut-off, so it measures real problem-solving instead of memorized answers.
It continuously pulls fresh problems from sites like LeetCode, AtCoder, and Codeforces and tags each with its release date.
A problem is a self-contained competitive-programming challenge, the kind posted weekly on LeetCode or Codeforces, and the model's code must pass all the hidden test cases on the first try. Because each problem is dated, a model can only be graded on ones it could not have seen during training.
LiveCodeBench is the benchmark built specifically to defeat cheating-by-memorization, which is why it is worth understanding even though it has a messy present. Its current state is itself the lesson: there is no single trustworthy up-to-date scoreboard, so the numbers quoted around the web come from three different aggregators that disagree with each other.
Pass@1, the share of problems solved on the first attempt, reported across difficulty tiers. The original board is a maintainer-run academic project on GitHub Pages; the catch is that it has not been updated since mid-2025.
Weak open models sit around 20 to 40 on Pass@1; on the frozen official board the strongest models reach the mid-80s on the full set and the low-70s on the recent-only window. No 2026 frontier model appears in the official data at all.
There is no trustworthy current leaderboard for LiveCodeBench, which is why this panel shows a dated anchor instead of a live top three. The official board is a JavaScript app last updated in mid-2025, with no 2026 model in it. Three separate aggregators each claim a different current top three (Artificial Analysis, vals.ai, and llm-stats, the last of which marks every entry self-reported), and they disagree because they use different problem sets and harnesses. The honest answer is that the benchmark's whole premise, dated problems, makes a frozen board unable to fairly rank newer models.
The benchmark creates the need to know. The catalog explains the ideas behind it:
No reliable current leaderboard. The most cited figure is a dated anchor:
Official LiveCodeBench v5 board (last updated AUG 2025, independently evaluated): on the full problem set the top Pass@1 scores are Kimi-k1.6-IOI-high at 86.02, Gemini-2.5-Pro at 84.66, and o1-2024-12-17 (High) at 83.18. The newest model in the official data is from early 2025; no 2026 frontier model appears.